A cluster attention-based multiple instance learning network for enhancing histopathological image interpretation

Background: Histopathological diagnosis involves examining abnormal architectural patterns and cellular-level changes. Whole slide images (WSIs) provide comprehensive digital representations of tissue samples, enabling detailed analysis and interpretation. Annotating the giga-pixel images remains labor-intensive, requiring experts to label abnormal patterns and cellular changes. To address this, Multiple Instance Learning (MIL), a promising weakly supervised approach, enables models to learn from limited annotations while preserving key histopathological features. Method: However, existing MIL-based methods may overlook potential semantic features, limiting their effectiveness. To overcome this limitation, we propose a novel Cluster-Aware Attention-based MIL (CAAMIL) architecture. This approach employs an advanced attention-based module integrated with a clustering method to enhance the interpretability of heterogeneous features. Our approach clusters architectural or cytologic features, making the groups interpretable at the cluster level and reflective of histopathological grades or prognostic indicators. Results: We demonstrated the efficacy of our model in both slide-level and patch-level classification as well as in interpreting tumor and mutation predictions. Experimental results show that our model achieves an AUC score of 0.96 for tumor detection at slide-level and 0.85 at patch-level, outperforming other state-of-the-art MIL-based methods. Conclusion: Our proposed CAAMIL architecture overcomes the limitations of existing MIL methods by effectively clustering features and providing interpretable results. The high accuracy and interpretability of our model make it a promising tool for histopathological diagnosis and tumor detection.